Search code examples
pythonamazon-web-servicesamazon-ec2deep-learningbert-language-model

How much faster is training on an AWS GPU enabled instance compared to training on 8-core CPU laptop?


I'm working on fine-tuning a BERT model that takes about 1 hour 20 minutes per epoch. I'm curious about how much more quickly I can expect this model to run after migrating to a g3s.xlarge. Looking for a percentage time savings / very ballpark estimate of runtime.


Solution

  • if your workload is suitable and things are set up well then 4 to 5 times faster

    see https://ieeexplore.ieee.org/document/8751930